AI Agents for Recruitment Agencies: Sourcing, Screening and Review Points

An AI agent can take the repeatable middle of a recruiter's desk: building a longlist from a job brief, scoring applicants against set criteria, drafting outreach, chasing replies and logging notes. A recruiter should review the criteria before the search, the first batch of matches, every rejection, and anything a candidate or client will read.

"Agent" covers a wide range in this market. Workable sells the Workable Agent, Loxo sells a "Loxo Agent Workforce", and Bullhorn calls its AI staff "digital workers". The label matters less than one question: after each step, who sees the result before the next step starts? Most agencies check the shortlist closely and never look at who the agent filtered out, which is where the damage hides.

Follow me on Instagram@sagnikteaches

What an agent can take off the desk, stage by stage

Treat the agent like a new resourcer on a trial period. It gets a job description with three columns: what it may do alone, what it must not do, and what a recruiter checks before the next stage begins. The right-hand column is the one to protect.

Connect on LinkedInSagnik Bhattacharya

Vendors split these stages differently, though the coverage is similar. Workable says its agent sources from 400M+ profiles and evaluates candidates against up to 14 criteria per job. Loxo says its Longlist agent screens up to 10,000 candidates in a search and its Shortlist agent ranks them across seven dimensions. Bullhorn's Amplify page lists digital-worker skills named Enrich, Match, Screen, Outreach and Present. You can read each in its own words: Workable's pricing page, Loxo's pricing page and Bullhorn Amplify.

Subscribe on YouTube@sagnikteaches

Four of the review points are non-negotiable in my view, so each gets a number. Gate 1 is the brief, gate 2 the first twenty matches, gate 3 the rejections, gate 4 anything outbound that states a fact. Together they cost a few hours per role, against far more for reading everything yourself, and the sums appear further down.

Gate 1: write the brief the agent will search on

An agent searches exactly what it is given. If the brief says "payroll experience" and nothing else, it will still rank two hundred people, and it will do so on whatever signals the text offers: job titles, software names, how recently someone left a job. A vague brief does not produce a vague answer. It produces a confident answer built on proxies.

Three rules keep the brief honest. First, cap the must-haves at three, because every extra one shrinks the pool and nudges the agent towards lookalike candidates. Second, get the client's deal-breakers in writing (notice period, work pattern, pay ceiling) and apply them as filters, not as score points. Third, list what must not be scored. Typical proxies are graduation year (a stand-in for age), employment gaps (caring, health, study), the candidate's name or spelling style, wording such as "digital native" or "recent graduate", and the university's name. Postcode is a proxy too: if the client sets a maximum commute, apply the client's rule, not the agent's guess.

Here is a filled-in brief for an illustrative role, written to be pasted into whatever job settings your agent offers.

ROLE: Payroll administrator (mid-sized employer, about 300 staff)

MUST-HAVE (three only)
1. Has run a full payroll cycle, weekly or monthly, end to end
2. Has used a payroll package (any named package counts)
3. Has met fixed monthly pay deadlines (evidence: ran payroll on time)

NICE-TO-HAVE (scored lower than must-haves)
- Has handled year-end reporting
- Has trained or supervised a junior colleague
- Has dealt with a pension or benefits provider

FILTERS (client confirmed in writing, applied before scoring)
- Notice period: three months or less
- Pay band: [client-approved range]
- Work pattern: three days on site

DO NOT SCORE
- Graduation year, employment gaps, name, photo, university
- Home area (the commute filter above is the client's rule)
- Phrases like "digital native", "energetic", "recent graduate"

Workable says its agent can be configured per account, per job and per candidate, so each job can carry its own brief. If the brief starts life as a client call, you can fill the job intake form from the call with AI first and then trim it to this shape.

Gate 2: read the first twenty matches before anyone is contacted

Run the search and stop at the first twenty. Nobody has been contacted; the agent has only built a list. Read the top twenty as if you were deciding whom to phone, and count the ones you would ring today. My starting rule is at least 12 of 20. Below that, the problem is almost always the brief rather than the talent pool: too many must-haves, a job title that means different things in different firms, or a software name the agent treats as a hard requirement. Edit the brief and rerun before spending more credits.

Job titles are the commonest culprit. "Payroll administrator", "payroll officer" and "payroll clerk" can describe one job or three. An agent that leans on titles will either miss the clerk who runs 300 payslips a month or rank an assistant who only keys in timesheets. When the first twenty look wrong, scan the reasons for the word the agent latched on to.

Two more checks belong here. Look for anyone with a do-not-contact flag, a current placement or a client restriction in your own records, because an outside profile pool will not know about your exclusions and not every agent reads those fields unless you set them up. And remember that passive candidates never applied to anything, so the first message is where your lawful basis for contacting them matters. The comparison of AI sourcing tools, their seat costs and consent rules covers that side.

Gate 3: check the rejection pile, not only the shortlist

Workable's homepage says strong fits auto-advance and clear "no" cases are routed out, with any call open to override in seconds. Loxo's pricing page says its agents are autonomous through the shortlist but never through the decision. Either design can work, provided one rule holds: the agent can sort, but a person sends the rejection. An agent's "no" is a prediction, and candidates who are wrongly filtered never appear anywhere you would notice.

The table shows what a small sample of screening output might look like, with the check a recruiter would make. The candidate IDs and findings are illustrative.

The first two errors would have stayed invisible if you only read the shortlist, because those candidates never reached it. A weekly sample of the rejections catches them. Pull 20 rejected candidates at random (or 10 per cent of the pile, if that is larger), reread each CV against the brief and count how many you would have progressed. If three or more of the 20 would have gone through, change the brief or move screening back to suggest-only for that job.

Before settling on that threshold, have a second recruiter screen the same 20 blind. Two careful humans disagree on some candidates too, and your baseline might be two in twenty. If so, three from the agent is ordinary and six is a problem. Also check that you can see a reason for every score. Workable's page says each action is logged and reversible, and Loxo shows actions in an Agent Impact Center. If a tool cannot tell you why it rejected someone, it should not be allowed to reject anyone.

Bullhorn's Screen skill is described as conducting, summarising and scoring candidate interviews, and its Verify skill as flagging possibly fraudulent records. For both, read the transcript or the flagged record yourself. A flag is a prompt to look, not a verdict.

Gate 4: outreach and chat run on facts you supply

The riskiest sentence an agent writes is the one that sounds plausible and states something nobody told it. Give the agent a fact sheet for each role and an instruction about what to do when the sheet runs out.

You are writing for [agency name] on behalf of a recruiter.
Use ONLY the facts below. If a candidate asks anything not covered,
reply "[Recruiter first name] will confirm that" and flag the chat.
Never state a salary, start date or working pattern that is not listed.
Never say a candidate is "a perfect match" or promise an interview.

FACTS
Role: payroll administrator, mid-sized employer
Pay: [client-approved range]
Working pattern: three days on site, two from home
Next step: ten-minute call with the recruiter
Opt-out: tell candidates they can ask not to be contacted again

Here is how a first draft can go wrong and what the recruiter's edit fixes. The draft reads: "Hi [first name], your profile is a perfect match for a payroll role paying up to $58,000 with full remote working. Reply YES and I'll book your interview this week!" It overclaims ("perfect match"), states a pay figure the client has not approved, promises remote working the spec never gave, and offers an interview before any screening. It also never says who is writing.

Hi [first name], I'm [your name] at [agency]. I'm recruiting a payroll administrator for a mid-sized employer and your payroll experience caught my eye. Pay and working pattern are in the brief, which I can send over. Would a ten-minute call this week suit you? If you'd rather not hear from me about roles, just say so and I'll remove you.

A subtler failure comes from chat. Picture an agent working from a spec line that says "flexible working". It tells eleven candidates the role is remote, and all eleven say they are interested. The mistake surfaces on day two, when a recruiter reads the transcript sample, and it costs a round of awkward calls rather than a client. Reading ten chats a week is enough to catch this kind of drift, and the fact sheet stops it recurring.

Chat is also where credits go fastest, so use it on people already worth speaking to. Check one handover behaviour as well: Workable says its agent stops the moment you engage a candidate. Test that on a dummy contact before you rely on it, because two voices messaging one candidate looks careless. For the written follow-through, the tutorials on automating candidate follow-up and on candidate summaries clients actually read go further.

What the credits cost, with a sum for one role

Agents are usually metered in credits, one per unit of work. Workable's pricing page lists three actions: 1 credit each time a candidate is screened against your ideal candidate profile, 2 each time a passive candidate is found and added to your pipeline, and 10 for each chat session in which a candidate gives at least one answer. Each paid plan includes 3,000 complimentary credits. Extra packs cost $0.12 a credit for 5,000 ($600), $0.10 for 10,000 ($1,000) and $0.095 for 50,000 ($4,750). Credits expire one year after purchase, they are non-refundable, and the free trial does not include the agent. The same page splits the agent into parts: a Job Brief Agent included on every plan, and Sourcing, Screening and Engagement agents sold as add-ons that draw on credits.

Here is an illustrative role: 240 applicants screened, 60 passive candidates added, and 45 candidates who answer in one chat session each, priced at the smallest pack's $0.12 a credit.

The free 3,000 credits cover roughly 3.7 roles of this size. The lesson is in the chat row: it is 450 of the 810 credits, more than screening and sourcing combined. Limit chat to 25 shortlisted or "maybe" candidates and the role falls to 610 credits, or $73.20. Check one wording point before you budget. The page charges per session, not per message, but it does not say what ends a session. If each of the 45 candidates comes back for a second session (to confirm availability a day later, say), chat doubles to 900 credits and the role comes to 1,260, or $151.20. Ask the vendor what opens and closes a session, and whether a candidate who never answers costs anything.

Set that against the time. Say a recruiter skims a CV in three minutes: 240 CVs is 12 hours. With the agent, the recruiter reads the top 25 at four minutes each (100 minutes), samples 24 rejections at two minutes each (48 minutes) and spends 20 minutes on the brief, which comes to 2.8 hours. That frees roughly 9 hours. At an internal cost of $35 an hour (use your own figure), the time saved is worth about $320 against $97 of credits for the whole role. The sum only holds if the reviews actually happen, which is why they are gates and not suggestions.

Other vendors meter differently. Loxo's pricing page lists a Core plan from $149 per user per month billed annually, with AI search but without the agents. Its Professional plan, priced by demo, includes 2,500 credits per user per month, pooled across the team and refreshed in full each month (unused credits do not roll over), with boosts to 5,000 (2x) or 12,500 (5x) per seat. The page says capacity figures are "figurative", and it does not say how many credits one full-market search uses. Ask for that number on your own typical role. Bullhorn publishes Starter at $99 and Core at $165 per user per month. Its Pro tier (Amplify AI features) and Max tier (Amplify Digital Workers) say "Talk to us" and show no credit figures. Put four questions to any vendor: what one credit buys, whether a re-screen after a brief change costs credits again, whether unused credits carry over, and whether the action log can be exported.

A bias check you can run in under an hour

You cannot prove an agent is fair, but you can find out whether it is moved by things it should ignore. The swap test does that. Write six fictional CVs for the role, three strong and three weak, so no real candidate's details are used. Make four variants of each: a different name, a 12-month career break added, the graduation year moved 15 years earlier, and a different home area. Run all 30 through the agent (the six originals plus 24 variants) and compare scores with the original.

Use a tolerance, for example 5 points on a 100-point scale. These results are illustrative:

A failure tells you what to write into the brief ("employment gaps are not scored; ignore dates of education") and what to retest. Run the swap test before go-live, after any change to the criteria, and whenever the vendor announces a model or scoring update, because those changes arrive without your say-so. Then keep the weekly rejection sample from gate 3 running; it catches what a fixed test set cannot.

If you collect equal-opportunities data lawfully, comparing pass rates by group each month is the stronger check. Ask a data-protection adviser before you analyse that data, and use the bias test checklist for small agencies for a fuller routine. If you recruit for employers in the EU, the AI Act lists recruitment among its high-risk uses (Annex III), with those obligations due from 2 December 2027, so building the audit trail now saves a scramble later.

Jobs the agent should never do alone

• Rejecting a candidate who reached a shortlist, an interview or a client submission. These people deserve a human voice.
• Anything about pay, offers or counter-offers. Money talk moves fast and mistakes become promises.
• References and background checks. An agent can chase the paperwork, but a recruiter reads the answers.
• Disability, health or other sensitive disclosures. Route the chat to a person at once.
• Complaints and data requests. Those need a named person and a record.
• Submitting a candidate to a client. Candidate consent comes first, and a recruiter signs off the final summary.

A two-week pilot for a six-recruiter agency

Imagine an agency of six recruiters filling payroll and accounts roles. It does not roll the agent out to every desk. It picks one role type with steady volume and one recruiter to own the pilot.

1. Day 1: choose one live job and write the brief (gate 1). Run the swap test on the six fictional CVs.
2. Days 2 to 4: run sourcing, read the first twenty (gate 2), and rerun if fewer than 12 are people you would call.
3. Days 5 to 9: screen inbound applicants with the agent set to suggest-only if the tool allows it. Sample 20 rejections daily for the first three days, then weekly.
4. Days 10 to 12: switch on outreach and chat for shortlisted candidates only, with the fact sheet. Read the first ten messages in full, then ten chat transcripts.
5. Day 14: fill in the scoreboard and decide to widen, hold or switch off.

If the numbers hold, widen the pilot to a second role type and add the recruiters who will review. If they do not, you have spent two weeks and one role's worth of credits finding that out, which beats finding it out through a client. When you compare tools, evaluating the AI features in your recruitment software gives a scoring sheet. If you would rather map the gates onto your own process with someone, that is what my AI implementation consultation covers, and you can start from the contact page.

Questions recruiters ask before switching an agent on

Do I have to tell candidates an AI agent is involved?

Tell them. It costs nothing, builds trust and avoids awkward surprises when a chat turns out not to be a person. If you recruit for employers in the EU, the AI Act's transparency rule expects people to be told they are dealing with an AI system. Add a line saying how to ask for a human, and ask an adviser about your own obligations.

Can ChatGPT or Claude act as the agent instead of an ATS add-on?

They can draft briefs, outreach and summaries, which covers several rows of the delegation table. They do not write into your ATS, log each action or enforce the review gates, so every gate becomes a manual habit. If candidate details go in, use a business plan, since the main business plans do not train on business content by default.

Does an agent replace a sourcer or resourcer?

It changes the job more than it removes it. First-pass reading and list-building shrink, while reviewing the no pile, calling candidates and managing the client grow. In the sums in this tutorial, screening review took under three hours against twelve hours of reading. Someone who knows the market still has to judge what the agent produces.

What if the agent's stated reasons do not match the CV?

Treat a mismatch as a reason to move that step back to suggest-only. Keep the examples, check whether the brief caused it, and rerun a sample. If the reasons still contradict the CV, raise it with the vendor, because a score you cannot explain to a candidate or a client is not safe to act on.

Further reads

• AI CV Screening for Recruitment Agencies: Setup and Safeguards — Setup and safeguards for the screening step in more detail.
• AI for Recruitment Agencies: Re-Engaging Your Candidate Database — Wake up your own database before paying for external sourcing.
• AI Boolean Search Generator for Recruiters: Prompts and Checks — Search-string prompts and checks if you still source by hand.
• AI Note Taker for Recruiters: Candidate Consent and Tool Choice — Candidate consent when an AI tool listens to your calls.
• Recruitment Agency KPIs: A Weekly Scorecard From Your CRM Export — Track pilot numbers next to placements and margin.
• How Much Time Does AI Save a Recruitment Agency per Role? — Compare your pilot minutes with realistic time savings per role.
• Is AI CV Screening Fair? Bias Checks Every Recruiter Should Run — More fairness checks aimed at the screening step.
• AI Talent Recruiting Software: Manatal, Loxo or Bullhorn Amplify? — Manatal from $15 a user, Loxo from $149, Bullhorn from $99 on its small-agency plans: what each costs a five-person agency and what the AI does.
• AI Staffing Best Practices: A Checklist for Placements and Candidate Data — A checklist for staffing agencies using AI in placements: who decides at each gate, what candidates are told, how long data is kept and what to log.
• How to Use Claude for Recruitment at a Small Agency: Specs and Outreach — Claude for job specs and candidate outreach at a small agency: working prompts, sample outputs and the rules that keep rejection decisions with a human.
• ICO AI Recruitment Report: A Checklist for UK Recruitment Agencies — An agency-facing checklist that turns the ICO recruitment findings into evidence, owners and review points.
• EU AI Act and Recruitment: What High-Risk Means for an Agency — A four-question test for whether your recruitment AI is high-risk, the dates that apply to a small agency, and what to do each quarter until 2027.
• Recruitment CRM AI: Candidate Records, Duplicates and Follow-Up — Use AI to organise candidate data and draft follow-up, while recruiters approve merges, corrections and contact.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Workable pricing, home and Workable AI pages, Loxo pricing page, Bullhorn Amplify and pricing pages, the EU AI Act Service Desk timeline and Annex III text, read on 1 November and re-checked on 4 November 2026.